A lightweight dual-branch spatial-frequency framework for joint deepfake detection and localization

In recent years, the rapid advancement of generative AI has made facial image manipulation increasingly accessible, raising serious concerns about AI-based attacks that can compromise face recognition systems or damage individual reputations, thereby necessitating forensic solutions that are accurate, interpretable, robust, and deployable in real-world conditions, particularly on resource-constrained devices. We propose a Dual-Branch Deepfake Detection (DBDD) model that jointly performs forgery detection and localization from spatial and frequency domains. The model not only predicts a binary label (real or fake) but also generates a forgery localization mask for the image, which serves as interpretable forensic evidence, supporting investigative workflows and increasing decision transparency. The spatial branch adopts a modified U 2 -Net backbone to capture multi-scale information. In parallel, the frequency branch employs encoder blocks built from Inception-based architectures combined with Convolutional Block Attention Module (CBAM) to encode frequency-domain inputs derived from Stationary Wavelet Transform (SWT) and Fast Fourier Transform (FFT). To enhance the ability to capture forgery cues in the frequency domain, we introduce a Hierarchical Adaptive Frequency Transform (HAFT) module that learns to emphasize informative regions before feeding the features into the encoder, instead of relying on fixed high-pass filters. In addition, a Multihead Fusion (MF) block is designed to effectively and comprehensively fuse the corresponding spatial and frequency features. Experiments on several benchmark datasets, including cross-dataset and cross-compression settings, demonstrate that the proposed model outperforms many existing methods on both detection and localization tasks, while using significantly fewer parameters and computational operations.